Articles On Babies and Bathwater - cs.ucf.edu

12 Pages·1995·2.32 MB·English

by  Patrick J. Hayes

Articles
On Babies and Bathwater
A Cautionary Tale
Patrick J. Hayes, Kenneth M. Ford, & Neil Agnew
n One should not throw out the baby with the Unfortunately, some newer nannies
bathwater, according to an old aphorism. Some haven’t been so careful, and the babies are in
popular recent positions in AI thinking have danger from their zealous ways. We will focus
done just this, we suggest, by rejecting the useful
on two nannies who seem to be close friends
idea of mental representations in their overen-
and often can be seen together—Situated
thusiastic zeal to correct some simplifications
Nanny (called SitNanny for short) and Radi-
and naïveties in the way traditional AI ideas have
cal Social Constructivist Nanny (known to
sometimes been understood. These “situated”
perspectives correctly emphasize that agents live her friends as RadNanny).1
in a social world, using their environments to
help guide their actions without needing to SitNanny Meets RepBaby
always plan their futures in detail; but they
incorrectly conclude that the very idea of mental SitNanny is fanatical about a certain kind of
representation is mistaken. This perspective has firmness. She just hates for things not to be
its intellectual roots in parts of recent sociologi-
firmly attached to the world. To be fair, this
cal thinking which reject the entire fabric of
obsession may be a natural consequence of
western science. We discuss these ideas and dis-
her having spent so much time in California.
putes in the form of an illustrated fable concern-
She believes that traditional ways of bring-
ing nannies and babies.
ing up baby are far, far too slack in this
regard. Babies shouldn’t think and plan so
much, she believes, but instead should be
kept firmly attached to the ground.2
A Cast of Characters
SitNanny declares (in fact, she preaches,
loudly and regularly) that all this old-fash-
O nce upon a time there were two hap- ioned cognitive stuff is just dirty bathwater,
py and healthy babies. We will call and she is going to throw it out. She believes
them Representation Baby (closely that it holds babies back, stunting their
related to Mind Baby and Person Baby) and development, and that a nursery based on
Science Baby (closely related to Reality Baby). such nonsense is an unhealthy place. She is
These babies were so charming and inspira- so fanatical about her views, however, that
tional that for a long time their nannies she is quite willing to throw out the baby
cared for them very well indeed. During this with the bathwater.
period it was generally the case that igno- Not long ago, Representation Baby (also
rance was pushed back and human dignity called RepBaby) was reflecting on a heretofore
increased. Nannies used honest, traditional happy childhood when rather suddenly he
methods of baby care which had evolved dur- found himself threatened by SitNanny, who
ing the years. Like many wise old folk, they adopts just this position.
were not always able to articulate good justi- SitNanny argues passionately with those
fications for their methods, but they worked, who suggest that keeping the baby might be a
and the healthy, happy babies were growing wise idea, going so far as to deny that there is
well and having lots of fun. a baby there at all. She argues further that her
Copyright © 1994, AAAI. 0738-4602-1994 / $2.00 FALL 1994 15
Articles
Situated Nanny Holding Things Down
16 AI MAGAZINE
Articles
task is simply to get the nursery clean in forth, the zero case is of interest. Such a ques-
preparation for other babies. tion certainly leads in a different direction
We think that SitNanny is at heart a good than the traditional AI work in planning and
person, and many of her suggested reforms focuses interest on different concerns. RepBa-
are valuable. The intellectual nursery did by can live with this. It provides at the very
need some tidying up. But she has been read- least a certain intellectual discipline. Babies,
ing political tracts (written by such people as even RepBabies, need to be set firmly on sol-
Dreyfus and Searle—the Berkeley Brothers) id ground quite often and (one might argue)
and has become something of a fanatic. should learn to walk before they try to play
Why would SitNanny want to threaten chess.
RepBaby, and what does it mean to be “situ- However, the term “situated” has become
ated”? There are several possible answers. identified with a much stronger and more
One emphasizes the fact that agents’ knowl- radical collection of assumptions, in particu-
edge of their world is incomplete, partial and lar an attack on the basic idea of knowledge
usually incorrect, but that they must never- representation. We believe that these assump-
theless act in that world, often quite prompt- tions are unwarranted and often based on
ly. This view leads to a vision that rejects the fundamental misunderstandings. In particu-
idea of planning in order to act. Planning, it lar, one can agree that AI should pay more
is said, makes the unjustified assumption attention to real-time activities that are inti-
that the planner has access to all relevant mately involved with the immediate environ-
facts, and it is a time-consuming activity that ment, without feeling a need to reject the
ignores the ticking of the real clock (Agre & entire framework of empirical science,
Chapman, 1987). It would be hard to play embrace radical social constructivism, or
baseball, for example, if one had to be always reject the idea of mental representation. Most
planning one’s actions. It seems clear that we of these more radical positions seem to arise
often, perhaps usually, simply start on a from reactions against what might be called
course of action in the blithe confidence that straw babies, versions of classical AI ideas The
we will be able to handle any difficulties that in fact are much too simplistic. SitNan-
intellectual
which may arise along the way. We just do it. ny seems to think that the AI concept of
And we often rely explicitly on the external “physical symbol system” amounts to a claim nursery did
world to guide our actions. If you are driving that programs, mental representations, and need some
in downtown San Francisco and want to get external texts are indistinguishable (e.g.,
tidying up.
to the airport, the best method is to find Clancey, 1993) and that the AI vision of
Highway 101 south, get in the left-hand lane, knowledge is of something rigid, unchange-
and stay in that lane through all subsequent able, and context-insensitive, a kind of “men-
junctions and turns. It will eventually tal toolkit” (Lave, 1988). In our view, all these
become the fast lane on the road to the air- positions misunderstand the essential con-
port. There is no need to plan a route: that tent of the physical symbol system hypothe-
left-hand lane will take you there. sis as it is most widely understood in AI. Par-
This disillusionment with the traditional ticular AI systems (often, ironically, those
AI emphasis on planning combines naturally that SitNanny was working on before she got
with an interest in how simple organisms get her new perspective) may make such simpli-
on. SitNanny has noticed that while insects, fying assumptions, but there is no reason
for example, have very impoverished cogni- why the general idea of symbolic representa-
tive abilities and never seem to plananything tion must be so limited, and indeed it tradi-
in the traditional AI sense of the term, there tionally has not been thought of in such a
are an awful lot of them around. Even quite limited way.
simple organisms can exhibit surprisingly The situationalists are attacking the very
complex behavior when placed in certain idea of knowledge representation—the
social and physical settings—a point made by notion that cognitive agents think about
Herb Simon (one of RepBaby’s grandfathers) their environments, in large part, by manipu-
years ago with his example of the ant travers- lating internal representations of the worlds
ing a beach of pebbles. they inhabit. Let us be frank: we think the
Further, it seems insightful to ask how representational hypothesis is a great idea.
much human behavior arises from, or might The reasons for being so positive are well
be simulated by, mechanisms that involve lit- documented, but we have two main justifica-
tle or no explicit representation of the exter- tions for our enthusiasm. First, it accounts for
nal environment. Even if one is concerned much that is otherwise completely puzzling
with representations and planning and so about how cognition could happen in the
FALL 1994 17
Articles
Representation Baby Is Worried about His Future
18 AI MAGAZINE
Articles
physical world; second, it allows experiments hoc justifications for action are not to be
and makes empirical predictions, which have identified with reasons for acting, apparently
so far largely been confirmed. But one needs believing that this confusion—which may
to understand the key word “representation” once have been a common one in the expert
in a sufficiently broad fashion. system field in which he worked—is endemic
Internal representations might not be con- to the representationalist perspective; but it is
sciously available to introspection, might uti- not. John Searle makes a similar mistake
lize ontological frameworks that are deter- when he concludes that anything represent-
mined by social or other contexts, might be ed must be equated with a conscious
involved with (and have their use involved thought, and any proposed mechanism of
in) social practices or any other kind of unconscious activity must therefore not
human activity, and might be involved in involve representations. Lucy Suchman
perceptual or motor skills at any cognitive (1987), a more subtle critic, also commits this
level. None of these are in any way at vari- mistake by drawing a sharp contrast between
ance with the representationalist hypothesis. cognitive science’s rigid, mechanical view of
RepBaby can play happily with these toys. plans and the weak resource view endorsed
What a representation means (not what by situated action theory. However these are
“representation” means, notice) is a complex not different views of a single idea, but two
question, but a meaning might be relative to distinct ideas: plans as programs or data
such things as goals, attitudes or purposes, structures, and plans as external maps or
and it need not presuppose that the environ- texts.
ment is uniquely determined. In fact, avail- SitNanny has been accused of this error
able semantic accounts of representation lan- before (Vera & Simon, 1993) and has protest-
guages—model theories—emphasize the ed that she does not mean to throw away all SitNanny
extent to which such relativity is probably notions of symbols or mental representation,
is unhappy
inevitable. but here she misses the point. What makes
SitNanny speaks with many voices, but a RepBaby so useful is that he makes physical with both
process of reconstruction and guesswork sug- symbols provide an explanation of mental
parts
gests to us that SitNanny understands repre- phenomena, and this explanatory role is
sentation in a different sense from that in what SitNanny rejects. Her new clean bath- of the
which it is used in the framework she rejects water will be scented with all manner of term
so vehemently. In fact, several of the new things: Gibsonian ideas of direct perception;
“knowledge
critics (some having noisily jumped ship) concepts from fringe neurology, sociology,
take representation to mean something like a ethnomethodology, and political theory; pre- represen-
text or a picture, an object consciously computational psychological theory; and
tation.”
manipulated by talking to oneself or visualiz- God knows what else. But the central idea of
ing internal imagery, and identical in nature physical symbol system will be purged from
to external representations such as writings it (Clancey, 1993; Greeno & Moore, 1993).
or diagrams. For example, Clancey (1993)
assumes that the use of a representation is Where’s the Beef?
always conscious and deliberate, and if a pro-
cess is “pre-linguistic” then representations When listening to SitNanny tell us that repre-
are not involved in it. But RepBaby was not sentations are not in the head, we often won-
brought up this way. It is at variance with the der, “Where’s the beef?” According to SitNan-
usual meaning of the term as used in the rep- ny, not only are there no mental
resentationalist position in cognitive science, representations, but individuals do not pos-
which hypothesizes internally represented sess knowledge. SitNanny is unhappy with
knowledge in a much broader sense, includ- both parts of the term “knowledge represen-
ing the computational modeling of uncon- tation.” Not only is the idea of representation
scious cognitive processes and processes of seen as misleading and false, but even the
social interaction. One might believe that notion of knowledge itself, as we understand
this broader sense is somehow inappropriate it, seems deeply suspect. Knowledge and
or incoherent, but such a position needs to meaning are seen as extra-personal and locat-
be argued, not simply asserted. Several of the ed in the community rather than in the head.
critics simply assume that calling something We need to get our terminology clear. Let’s
a representation entails that it has this exter- take a simple example. A plumber and a cus-
nal, perceived character. We will call this mis- tomer who knows nothing about plumbing
take the “textual fallacy.” are together in a kitchen. Where is the
Clancey (correctly) emphasizes that post knowledge of plumbing? RepBaby has an
FALL 1994 19
Articles
almost childishly simple answer to this ques- and places the meanings not, as it were, dis-
tion: It’s in the plumber’s head! Is RepBaby tributed among the heads of the members of
right? Well, let us try to converge on the the society, but in the society itself. Going
locus of plumbing knowledge. If someone one step beyond Jung, we have to imagine
were to ask a question about hot-water cylin- something like an invisible collective con-
ders, it would be correct to tell him that the sciousness—what Harry Collins calls a “col-
answer could be found in the kitchen, so it lectivity”—in which the meanings of our
seems correct to say that the knowledge is in individual beliefs are located. The old AI idea
the kitchen. of common sense might fit easily here: what
Where in the kitchen is it, then? Perform a more natural place to locate the sense com-
simple experiment by removing the plumber mon to the members of a society than in that
from the kitchen. Now nobody in the society itself?
kitchen knows much about plumbing, and it We are quite sympathetic to this view,
would be wrong to say that that was where properly understood. But notice the differ-
the answer could be found. It seems clear ence between this idea and the earlier one.
that the plumber’s knowledge of plumbing is This idea says that meanings of mental repre-
firmly attached to the plumber. Perhaps some sentations are socially determined; the other
kind of invisible mental leash would explain claims that the representational tokens them-
We believe its attachment to him, but we suggest that a selves aren’t in the head or that representa-
simpler and more plausible theory of this tional tokens can have only an external,
that
phenomenon might be that the plumber’s social existence, or even that there isn’t any
SitNanny’s knowledge of plumbing is in his head and representation at all. This conflict is like the
therefore moves with him, much as his kid- difference between insisting that RepBaby
attack on
neys and his toenails do. wear fashionable clothes and claiming that
RepBaby is Memories provide an even more vivid illus- there isn’t anyone there to be dressed.
unwarranted tration of beliefs being in the head. We all Suppose the plumber in the kitchen real-
have memories that no one else shares. Sure- izes he has forgotten his flux and knows that
and often
ly, these are not located anywhere but in our a soldered joint might therefore be unreli-
based on own heads (or at most our own bodies). A able, so he contemplates using a compression
memory can be damaged by banging a head, fitting in spite of its extra cost. Perhaps
fundamental
but never by stubbing a toe or digging a hole. indeed a proper account of what these terms
misunder- Notice this is not to claim that memories mean might involve a description of the
are veridical or that they are anything as sim- plumbing community and his relationship to
standings.
ple as a replaying of earlier recordings. They it. Any experience, including this one, may
might be the result of some complex and cre- change and enrich his knowledge of joining
RepBaby ative process of reconstruction with which we pipes. But these are mental tools that he
all constantly rewrite our past. And it might be uses—or, more properly, mental mechanisms
is correct to say that we are almost never doing out of which he is formed—and when he
worth just one thing, that our actions are always leaves the kitchen, they go with him.
embedded in some social context; although In summary, the central concept of mental
saving.
we think this fact is less significant than Sit- representation is more robust than realized
Nanny does. But whatever is going on when by SitNanny. We believe that SitNanny’s
we think or when we are recalling something attack on RepBaby is unwarranted and often
from our past, it is happening inside us, not based on fundamental misunderstandings.
somehow out there in the society. RepBaby is worth saving.
There is, however, a rather different way to
understand SitNanny’s social-meaning claim. “Hey, Hey, Ho, Ho,
One might argue that although there are Science Baby Must Go”
things in the head we call representations,
any account of the meaning of these things Now we come to Science Baby—an extremely
(and hence any account of why they should robust and healthy lad. SitNanny doesn’t
be regarded as representations) must be bear Science Baby much harm. But RadNan-
inherently social. Our beliefs have meaning ny, one of SitNanny’s mentors, is much more
only by virtue of their role in a society of dangerous—RadNanny belongs to a cult.
which we are part. Wittgenstein argued that Remarkably, even this most useful baby is
the idea of a language spoken by a single per- under attack by RadNanny and the members
son is incoherent: languages are systems of of her cult.
communication between agents. The socio- Like an anthropologist studying a primi-
logical stance takes this idea rather further tive tribe, RadNanny (and her band of sociol-
20 AI MAGAZINE
Articles
Science Baby Doing His Sums
FALL 1994 21
Articles
The Notorious RadNanny Looking for Babies
22 AI MAGAZINE
Articles
ogists) has moved into scientific laboratories SitNanny often enthusiastically endorses
to peer over the shoulders of scientists. Rad- this radical social constructivism because it
Nanny has noticed that the behavior of real can be seen as locating knowledge in the
scientists typically does not follow the community rather than in the head and so
abstract ideal of disinterested pursuit of truth. can be taken to support a position that denies
However RadNanny draws the startling and a central role to internal representations. We
extreme conclusion that all science is arbi- sometimes wonder if SitNanny really under-
trary and that reality is merely a construction stands the full import of RadNanny’s doc-
of a social game. trines.
This anti-epistemic and also anti-scientific It is important to notice that it is not just
stance is clear in statements such as “the nat- Science Baby that is under attack here, but
ural world has a small or nonexistent role in the whole notion of an objective reality. Rad-
the construction of scientific knowledge” Nanny’s ideas are a sort of universal baby sol-
(Collins 1981, p. 3) and “the fashioning of vent. However, we are not too worried about
normative models of thinking from particu- Reality Baby. Even RadNanny will duck if one
lar, ‘scientific,’ culturally valued, named bod- throws a brick at her.
ies of knowledge is a cultural act” (Lave 1988,
p. 172). RadNanny claims that since we have Conclusion
no direct access to “reality,” science (like reli-
gion, politics, and literature) merely reflects These new nannies have a lot of valuable
the myth-making tendencies of the human reforms to make. We aren’t trying to argue
tribe, with one myth being no more reality- simply for conservatism. Sometimes the old-
based than another. fashioned nannies were (and still are) using
On this view, science should not be grant- dirty bathwater. One does need to keep Rep-
ed any kind of relative authority or have any Baby under control—if you let him, he will
particular success acknowledged. It is just one try to get hold of everything.
cultural tradition: Zen Buddhism, medieval SitNanny is right to emphasize that a lot
Christianity, or even the semi-coherent cyni- can be done without involving detailed,
cism of Beavis and Butthead are all perfectly explicit knowledge of it. Sometimes the best
valid alternatives. In fact, sometimes scientif- plan is just to follow the road. But this
ic rationality is seen as morally or politically musn’t be taken too far. It is really an obser-
inferior to Beavis and Butthead: “Logic, as vation about what kind of information
Marx has it, is the money of the mind, and should be represented rather than a rejection
no matter how dialectical, it always expresses of the idea of knowledge representation itself.
a reified and alienated mediation of man and In any case, the question of how much plan-
reality” (Warren, 1984 p. 50, cited in Lave, ning goes on is essentially an empirical one.
1988 p. 173). We have criticized this cultural Something clever happens when a good base-
relativism elsewhere (Agnew, Ford, & Hayes, ball fielder begins to run toward a catch
1994), so here we will be brief. before the bat has made contact with the
By any reasonable measure, Science Baby ball. Maybe this isn’t planning in a very sim-
has been spectacularly successful. Perhaps no plistic sense, but it isn’t just following the left
single scientist operates according to the ide- lane either.
al rules, as RadNanny delights in telling us, Even RadNanny has some things of value
but the approximate global effect is to pro- to give us. A certain kind of naïve realism is
duce knowledge that reality (the hidden- wrong, and human scientists sometimes do
hand editor) has had an opportunity to grip all sorts of strange human things. But if the
as firmly as it can be made to. radical constructivists would just take the
RadNanny claims that the natural world time to actually look at what the knowledge
plays little or no part in the beliefs of scien- representation idea means, they would see
tists, and that logic is only a tool used by one that it has already passed that level of
class to enslave others; but what can be the naïveté. If someone’s perceptions are repre-
basis for her beliefs other than empirical sentations, then of course they aren’t in
observation of the natural world (which, of direct touch with the world (whatever that
course, includes the scientists she studies), might mean). RepBaby is a pragmatic con-
and what logic does she use to explain her structivist, but he gets on quite happily with
ideas? Are RadNanny’s sociological friends Reality Baby. Science Baby likes both of them.
somehow issued special glasses that allow RadNanny, SitNanny, and others of height-
them to see more clearly than other scien- ened social awareness have been doing some
tists? new and exciting work that can rinse away a
FALL 1994 23
Articles
Reality Baby in a Relativist Stew
24 AI MAGAZINE